Repository navigation
v1.1.0
Sibyl 1.1.0
Released: 2026-07-24
This release turns Sibyl's memory into a usage-aware loop: memories you cite and revisit survive longer, wrong ones fade faster, and every write returns a durable receipt you can inspect and correct. It adds team-scoped memory, portable OKF graph exports, an accurate multi-step retrieval mode, and hardened SurrealDB 3.2.3 runtime pins across Docker and Helm.
🌟 Highlights
⚡ Usage-aware forgetting
Sibyl now tracks how each memory is used and feeds that signal into retention. services/usage.py records three signals: EXPOSURE (returned in results), CITATION (actively referenced), and MISLED (negative feedback). usage_retention_multiplier() scales a memory's decay half-life from 0.1x to 4.0x, so a heavily cited decision survives roughly four times longer than a random capture while items flagged misleading age out fast. Nightly consolidation orders decay candidates by retrieval count first, so memories you actively use are no longer archived by surprise.
🤖 Team memory control plane
A new TEAM memory scope sits between PROJECT and organization-wide sharing. The /teams route (apps/api/src/sibyl/api/routes/teams.py) adds nine endpoints for team CRUD, membership, and project links, all gated to OWNER/ADMIN org roles. Team read/write/reflect is authorized by verified membership against accessible_teams, and crossing out of team scope still requires an explicit SHARE promotion rather than silent leakage.
✨ Portable OKF graph export
sibyld export okf projects a graph archive into an Open Knowledge Format v0.1 markdown tree (export/okf.py): entities/, relationships/, episodes/, and mentions/ directories plus index.md and log.md. Exports are deterministic and round-trip verifiable, so you can take your knowledge graph out of Sibyl into plain, diff-friendly files.
✨ Accurate retrieval mode with anchor rescue
The context evidence API gains a retrieval_mode flag. fast (default) preserves existing single-query behavior; accurate plans up to three supplemental score-blind queries across evidence facets and fuses results with reciprocal rank fusion. Quoted or backticked text in a query is now guaranteed a slot in the evidence window via anchor rescue (query_anchors.py), and a deterministic refinement planner derives follow-up searches without extra LLM latency.
♻️ Correct and blame your memory
Write operations return a MutationReceipt with revision and affected-record tracking for idempotent replay. New sibyl correct and sibyl blame verbs let you mark memories wrong, stale, duplicate, or superseded, revise canonical content, and inspect the full correction and audit lineage. A new operational experience write path (/memory/experience) captures goal-directed observation sequences and distills them into typed workflow, fact, and gotcha notes.
Memory and retrieval
- Added
MemoryUsageSignal,MemoryUsageEvent, andMemoryUsageStampinservices/usage.py, with deduplicated event writes tomemory_usage_eventsand counter backfill during migration. - Exposure stamping (
tools/usage_exposure.py) annotates search and context results with stable cite IDs; internal synthesis and fallback paths opt out to avoid polluting recency. - Citation recording (
tools/usage_citation.py) accepts positive citations andMISLEDfeedback, returning coverage reports for which IDs were stamped. - Score-blind query planner (
retrieval/query_planning.py) plans up to four facet-labeled queries and rejects web-search markers (site:,filetype:,http://) so private-memory retrieval never escapes to the public web. - Operational evidence composition (
retrieval/operational_evidence.py) reserves a typed-note lane and selects bounded contiguous observation windows via manifest-validated source inventory. - Async note distillation (
ai/operational_distillation.py) produces imperative workflow recipes, concrete environment facts, and pitfall gotchas from captured sessions, named against real UI labels. include_retrieval_diagnosticssurfaces planner status, per-query counts, composition mode, and provider usage/cost for auditability.
Trust, evaluation, and provenance
- Added trust gates under
tools/trust/:write_path_integrity_gate.py(zero-budget contracts for hallucinated, self-referential, and low-signal writes),forgetting_gate.py,usage_loop_gate.py,okf_export_gate.py, anddoc_claim_gate.py. - Added dogfood receipt collection (
dogfood_receipts.py) that pulls live-deployment evidence, verifies version and image-digest parity, and checks required source-commit coverage. - Expanded LongMemEval V2 evaluation lanes: official protected paid runs, live smoke, composition replay, reader holdout planning and reporting, causal ablation, and validation slices.
- Added runtime provenance (
runtime_provenance.py) so the health endpoint reports the deployed version and commit.
Developer experience
sibylCLI verbs converged: capture, note, correct, blame, cite, and memory-promote now share consistent surfaces and provenance metadata (capture_mode,capture_surface,remember_kind).feat(cli): route to the right server by directoryselects the correct server based on working directory.- Richer CLI search previews and prioritized agent memory in context assembly.
- Skill packs tightened from two weeks of mined agent friction; added a compact agent
contract.mdskill pack. - Web UI: organizations page is now org-admin only, added a teams settings page, surfaced graph fetch errors, and fixed token drift and ID truncation.
Documentation
- README and package docs refreshed for the 1.1 surface; skills reframed as cross-agent and built into the CLI.
- Self-hosting docs now lead with local self-host as the default, enterprise as opt-in; added
docs/guide/self-hosting.mdanddocs/cli/team.md. - Added the post-1.0 roadmap (v1.1 to v1.3) and removed superseded historical planning docs.
Breaking Changes
- SurrealDB pinned to v3.2.3. Docker runtime, local-surreal, and Helm
surrealdbchart (values.yamlimage.tag: v3.2.3) now target 3.2.3. Fulltext query arms were made conjunctive-safe (fix(retrieval): make fulltext arms conjunctive-safe for SurrealDB 3.2); baseline parity passes on both 3.2.3 and 3.1.0. - Helm requires a stable settings encryption key. Deployments without a pre-configured encryption key will fail render validation. Set the key before upgrading.
- Team control-plane endpoints are org-admin only.
/teamsoperations requireOWNERorADMIN; member-level clients cannot modify teams.
Upgrade Notes
- Set a stable settings encryption key in your Helm values before deploying 1.1.0.
- Update SurrealDB to v3.2.3 to match the pinned runtime; the fulltext fix keeps 3.1.0 working if you cannot move yet.
- Retrieval defaults are unchanged:
retrieval_modestaysfast. Opt intoaccurateper request to enable multi-step planning. - No action needed for usage-aware forgetting; counters backfill during migration and decay ordering applies on the next consolidation pass.
- Docker images now publish to both GHCR and Docker Hub with enforced manifest-digest parity; ensure both registries are configured if you mirror releases.